I'm a postdoctoral research fellow in the Laboratory for Systems Pharmacology and Nirmal Lab at Harvard Medical School/Brigham and Women's Hospital. I combine single-cell transcriptomics, imaging, and deep learning to better understand cancer.
Personal Website: https://www.jost.engineer
Email: ty.jost@gmail.com
LinkedIn: www.linkedin.com/in/tyler-jost
If you are interested in trained models, imaging data, or more details, please reach out!
clusterCleaver is a computational pipeline for analyzing cluster data in single-cell RNA sequencing (scRNAseq). You can check out our paper here. This is a scanpy-compatible package which leverages the Earth Mover's Distance (EMD) to find genes which can be used to distinguish between clusters of cells. We also validated that our method worked, separating two breast cancer transcriptomic subpopulations. I generated the algorithm and performed the sequencing analysis.
UMAP and gene expression histograms for a top MDA-MB-231 cell line marker, ESAM.
Repo: https://github.com/brocklab/clusterCleaver
This week in music is a personal project I started to let myself know about live bands playing near me. I first scrape upcoming concerts near me. However, these are often presented like:
Australian punk night presents: THE CHATS / COSMIC PSYCHOS / THE SCHIZOPHONICS
This is pretty hard to parse with just a simple regex. So I finetuned a BERT LLM to extract these names. I then used the Spotify API to add songs from these artists into a playlist for the upcoming week. Originally it was just for Austin, TX but I've now expanded it to Boston and hopefully can do it for more cities.
Repo: https://github.com/TylerJost/twia
Cell morphology work often focuses on shape and texture, and typically identification is not done on subpopulations within a cell line but instead on other features such as metastatic variability, gene deletions, etc. We wanted to know if cells with transcriptomic changes within a cell line could also be identified using basic phase contrast imaging. To do so, we used our isolated populations found from the statistical approach mentioned above. We labeled each subpopulations and used a convolutional neural network to identify subpopulation. We found that, despite being relatively unperturbed, intra cell line populations are identifiable through computer vision. Additionally, we found that including neighborhood interactions instead of just shape and texture (possible primarily due to our use of deep learning) allowed us to improve our classification abilities. You can check out our preprint here. It's in press and will soon be published.
Repo: https://github.com/brocklab/transcriptomicClusterMorph